Objective approach to diagnosing attention deficit hyperactivity disorder by using pixel subtraction and machine learning classification of outpatient consultation videos

Background Attention deficit hyperactivity disorder (ADHD) is a common childhood neurodevelopmental disorder, affecting between 5% and 7% of school-age children. ADHD is typically characterized by persistent patterns of inattention or hyperactivity-impulsivity, and it is diagnosed on the basis of the criteria outlined in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, through subjective observations and information provided by parents and teachers. Diagnosing ADHD in children is challenging, despite several assessment tools, such as the Swanson, Nolan, and Pelham questionnaire, being widely available. Such scales provide only a subjective understanding of the disorder. In this study, we employed video pixel subtraction and machine learning classification to objectively categorize 85 participants (43 with a diagnosis of ADHD and 42 without) into an ADHD group or a non-ADHD group by quantifying their movements. Methods We employed pixel subtraction movement quantization by analyzing movement features in videos of patients in outpatient consultation rooms. Pixel subtraction is a technique in which the number of pixels in one frame is subtracted from that in another frame to detect changes between the two frames. A difference between the pixel values indicates the presence of movement. In the current study, the patients' subtracted image sequences were characterized using three movement feature values: mean, variance, and Shannon entropy value. A classification analysis based on six machine learning models was performed to compare the performance indices and the discriminatory power of various features. Results The results revealed that compared with the non-ADHD group, the ADHD group had significantly larger values for all movement features. Notably, the Shannon entropy values were 2.38 +/- 0.59 and 1.0 +/- 0.38 in the ADHD and non-ADHD groups, respectively (P < 0.0001). The Random Forest machine learning classification model achieved the most favorable results, with an accuracy of 90.24%, sensitivity of 88.85%, specificity of 91.75%, and area under the curve of 93.87%. Conclusion Our pixel subtraction and machine learning classification approach is an objective and practical method that can aid to clinical decisions regarding ADHD diagnosis.